
We have now started and graded 659 MCP servers in an isolated sandbox with no keys and no accounts. The average grade is 85 out of 100. That is a better number than we expected when we began.
Where points are won
Each server is scored out of 20 on five checks. Averaged across everything we graded, here is how they came out:
- Package safety: 19.1 out of 20
- Maintenance: 18.5 out of 20
- Tool descriptions: 18.2 out of 20
- Runs: 18.0 out of 20
- Adoption: 11.2 out of 20
In plain terms, most of these servers install, start and describe their tools well enough for an AI to use them. The code is mostly in decent shape.
Where points are lost
Adoption is the weak spot by a wide margin. The typical graded server gets about 1,500 npm downloads a week. Only 82 of the 659 pass 10,000, and 223 are under 1,000. That matters because a lightly used server has fewer people noticing bugs, fewer fixes, and less proof that it holds up outside the author's own setup.
What we would do with this
If you are choosing a server, a high grade tells you it starts and is maintained. Check adoption separately, and prefer the ones other people already rely on. If you publish a server, the cheapest improvement is not more features. It is making it easy to find, install and trust.
These numbers come from our own automated tests and will move as we test more servers. See how we grade on the methodology page.
Get the weekly insights by email
One email a week with what our tests found. Unsubscribe any time.
Want this server set up for your team?
ConnectATD builds AI automation and wires up MCP servers like this one. Tell us what you are trying to do.
Talk to ConnectATD